Bioinformatics 3 V 4 – Weak Indicators and Communities

1. Bioinformatics 3V 4 – Weak Indicators and Communities Mon, Oct 28, 2013

2. TAP Y2H annotated: septin complex HMS-PCI Noisy Data — Clear Statements? For yeast: ~ 6000 proteins → ~18 million potential interactions rough estimates: ≤ 100000 interactions occur → 1 true positive for 200 potential candidates = 0.5% → decisive experiment must have accuracy << 0.5% false positives Different experiments detect different interactions For yeast: 80000 interactions known, only 2400 found in > 1 experiment Y2H: → many false positives (up to 50% errors) Co-expression: → gives indications at best Combine weak indicators = ??? see: von Mering (2002)

3. P(A) = P(B) = P(B | A) = P(A | B) = prior probability (marginal prob.) for "A" → no prior knowledge about A prior probability for "B" → normalizing constant conditional probability for "B given A" posterior probability for "A given B" P(A) P(A ⋂ B) P(B) Conditional Probabilities Joint probability for "A and B": Solve for conditional probability for "A when B is true" → Bayes' Theorem: → Use information about B to improve knowledge about A

4. P(A) P(A ⋂ B) P(B) What are the Odds? Express Bayes theorem in terms of odds: • Also Consider case "A does not apply": • odds for A when we know about B (we will interpret B as information or features): posterior odds for A likelihood ratio prior odds for A Λ(A | B) → by how much does our knowledge about A improve?

5. 2 types of Bayesian Networks Encode conditional dependencies between evidences = "A depends on B" with the conditional probability P(A | B) Evidence nodes can have a variety of types: numbers, categories, … (1) Naive Bayesian network → independent odds (2) Fully connected Bayesian network → table of joint odds

6. Bayesian Analysis of Complexes Science 302 (2003) 449

7. Improving the Odds Is a given protein pair AB a complex (from all that we know)? likelihood ratio: improvement of the odds when we know about features f1, f2, … prior odds for a random pair AB to be a complex Idea: determine from known complexes and use for prediction of new complexes estimate (somehow) Features used by Jansen et al (2003): • 4 experimental data sets of complexes • mRNA co-expression profiles • biological functions annotated to the proteins (GO, MIPS) • essentiality for the cell

8. Gold Standard Sets To determine → use two data sets with known features f1, f2, … for training Requirements for training data: i) independent of the data serving as evidence ii) large enough for good statistics iii) free of systematic bias Gold Standard Positive Set (GP): 8250 complexes from the hand-curated MIPS catalog of protein complexes (MIPS stands for Munich Information Center for Protein Sequences) Gold Standard Negative Set (GN): 2708746 (non-)complexes formed by proteins from different cellular compartments (assuming that such protein pairs likely do not interact)

9. Prior Odds Jansen et al: • estimated ≥ 30000 existing complexes in yeast • 18 Mio. possible complexes →P(Complex) ≈ 1/600 → Oprior = 1/600 → The odds are 600 : 1 against picking a complex at random → expect 50% good hits (TP > FP) with  ≈ 600 Note: Oprior is mostly an educated guess

10. 0.19 0.36 1114 2150 = 0,5 = 0,518 Essentiality Test whether both proteins are essential (E) for the cell or not (N) → for protein complexes, EE or NN should occur more often pos/neg: # of gold standard positives/ negatives with essentiality information possible values of the feature overlap of gold standard sets with feature values probabilities for each feature value likelihood ratios

11. mRNA Co-Expression Publicly available expression data from • the Rosetta compendium • the yeast cell cycle ) Correlation between the data sets → use principal component Jansen et al, Science 302 (2003) 449

12. Biological Function Use MIPS function catalog and Gene Ontology function annotations • determine functional class shared by the two proteins; small values (1-9) Indicate highest MIPS function or GO BP similarity • count how many of the 18 Mio potential pairs share this classification Jansen et al, Science 302 (2003) 449

13. Experimental Data Sets In vivo pull-down: Gavin et al, Nature415 (2002) 141 Ho et al, Nature415 (2002) 180 31304 pairs 25333 pairs 981 pairs 4393 pairs HT-Y2H: Uetz et al, Nature403 (2000) 623 Ito et al, PNAS98 (2001) 4569 4 experiments on overlapping PP pairs → 24 = 16 categories — fully connected Bayes network Jansen et al, Science 302 (2003) 449

14. Statistical Uncertainties 1) L(1111) < L(1001) statistical uncertainty: Overlap with all experiments is smaller → larger uncertainty 2) L(1110) = NaN? Use conservative lower bound → assume 1 overlap with GN →L(1110) ≥ 1970 Jansen et al, Science 302 (2003) 449

15. Overview Jansen et al, Science 302 (2003) 449

16. Performance of complex prediction Re-classify Gold standard complexes: Ratio of true positives to false positives → None of the evidences alone was enough Jansen et al, Science 302 (2003) 449

17. Coverage Predicted set covers 27% of the GP Jansen et al, Science 302 (2003) 449

18. Verification of Predicted Complexes Compare predicted complexes with available experimental evidence and directed new TAP-tag experiments → use directed experiments to verify new predictions (more efficient) Jansen et al, Science 302 (2003) 449

19. Consider Connectivity Only take proteins e.g. with ≥ 20 links → This preserves links inside the complex, filter false-positive links to heterogeneous groups outside the complex Jansen et al, Science 302 (2003) 449

20. Follow-up work: PrePPI (2012) Given a pair of query proteins that potentially interact (QA, QB), representative structures for the individual subunits (MA, MB) are taken from the PDB, where available, or from homology model databases. For each subunit we find both close and remote structural neighbours. A ‘template’ for the interaction exists whenever a PDB or PQS structure contains a pair of interacting chains (for example, NA1–NB3) that are structural neighbours of MA and MB, respectively. A model is constructed by superposing the individual subunits, MA and MB, on their corresponding structural neighbours, NA1 and NB3. We assign 5 empirical-structure-based scores to each interaction model and then calculate a likelihood for each model to represent a true interaction by combining these scores using a Bayesian network trained on the HC and the N interaction reference sets. We finally combine the structure-derived score (SM) with non-structural evidence associated with the query proteins (for example, co-expression, functional similarity) using a naive Bayesian classifier. Zhang et al, Nature (2012) 490, 556–560

21. Results of PrePPI Receiver-operator characteristics (ROC) for predicted yeast complexes. Examined features: - structural modeling (SM), - GO similarity, - protein essentiality (ES) relationship, - MIPS similarity, - co‐expression (CE), - phylogenetic profile (PP) similarity. Also listed are 2 combinations: - NS for the integration of all non‐structure clues, i.e. GO, ES, MIPS, CE, and PP, and - PrePPI for all structural and non‐structure clues). This gave 30.000 high-confidence PP interactions for yeast and 300.000 for human. Jansen et al, Science 302 (2003) 449

22. Summary: Bayesian Analysis Combination of weak features yields powerful predictions • boosts odds via Bayes' theorem • Gold standard sets for training the likelihood ratios Bayes vs. other machine learning techniques: (voting, unions, SVM, neuronal networks, decision trees, …) →arbitrary types of data can be combined → weight data according to their reliability → include conditional relations between evidences → easily accommodates missing data (e.g., zero overlap with GN) →transparent procedure → predictions easy to interpret

23. Connected Regions Observation: more interactions inside a complex than to the outside → how can one identify highly connected regions in a network? 1) Fully connected region: Clique clique := G' = (V', E' = V'(2)) Problems with cliques: • finding cliques is NP-hard (but "works" O(N2) for the sparsely connected biological networks) • biological protein complexes are not always fully connected

24. Communities subset of vertices, for which the internal connectivity is denser than to the outside Community := Aim: map network onto tree that reflects the community structure ??? <=> Radicchi et al, PNAS101 (2004) 2658:

25. Hierarchical Clustering 1) Assign a weight Wij to each pair of vertices i, j that measures how "closely related" these two vertices are. 2) Iteratively add edges between pairs of nodes with decreasing Wij Measures for Wij: 1) Number of vertex-independent paths between vertices i and j (vertex-independent paths between i and j: no shared vertex except i and j) Menger (1927): the number of vertex-independent paths equals the number of vertices that have to be removed to cut all paths between i and j → measure for network robustness 2) Number of edge-independent paths between i and j 3) Total number of pathsL between i and j but L = 0 or ∞ → weight paths with their length αL with α < 1 Problem: vertices with a single link are separated from the communities

26. Vertex Betweenness Freeman (1927): count on how many shortest paths a vertex is visited For a graph G = (V, E) with |V| = n Betweenness for vertex ν: Alternative: edge betweenness → to how many shortest paths does this edge belong

27. Girvan-Newman Algorithm Girvan, Newman, PNAS99 (2002) 7821: For a graph G = (V, E) with |V| = n, |E| = m 1) Calculate betweenness for all m edges (takes O(mn) time) 2) Remove edge with highest betweenness 3) Recalculate betweenness for all affected nodes 4) Repeat from 2) until no more edge is left (at most n iterations) 5) Build up tree from V by reinserting vertices in reverse order Works well, but slow: O(mn2) ≈ O(n3) for scale-free networks (|E| = 2 |V|) Reason for complexity: compute shortest paths (n2) for m edges → recalculating a global property is expensive for larger networks

28. Zachary's Karate Club • observed friendship relations of 34 members over two years • correlate fractions at break-up with calculated communities with edge betweenness: with number of edge-independent paths: administrator's faction instructor's faction Girvan, Newman, PNAS99 (2002) 7821

29. Collaboration Network The largest component of the Santa Fe Institute collaboration network, with the primary divisions detected by the GN algorithm indicated by different vertex shapes. Edge: two authors have co-authored a joint paper. Girvan, Newman, PNAS99 (2002) 7821

30. Determining Communities Faster Radicchi et al, PNAS101 (2004) 2658: Determine edge weights via edge-clustering coefficient → local measure → much faster, esp. for large networks Modified edge-clustering coefficient: → fraction of potential triangles with edge between i and j k = 4 k = 5 Here, zi,j(3) is the number of triangles, kiand kjare the degrees of nodes i and j. Note: "+ 1" to remove degeneracy for zi,j(3) = 0 C(3) = (2+1) / 3 = 1

31. Performance Instead of triangles: cycles of higher order g → continuous transition to a global measure Radicchi et al-algorithm: O(N2) for large networks Radicchi et al, PNAS101 (2004) 2658:

32. Comparison of algorithms Data set: football teams from US colleges; different symbols = different conferences, teams played ca. 7 intraconference games and 4 inter-conference games in 2000 season. Girven-Newman algorithm Radicchi with g = 4 → very similar communities

33. Strong Communities "Community := subgraph with more interactions inside than to the outside" A subgraph V is a community in a… …strong sense when: …weak sense when: → Check every node individually → allow for borderline nodes Radicchi et al, PNAS101 (2004) 2658 • Σ kin = 2, Σ kout = 1 {kin, kout} = {1,1}, {1,0} → community in a weak sense • Σ kin = 10, Σ kout = 2 {kin, kout} = {2,1}, {2, 0}, {3, 1}, {2,0}, {1,0} → community in a strong and weak sense

34. Summary What you learned today: • how to combine a set of noisy evidences into a powerful prediction tool → Bayes analysis • how to find communities in a network efficiently → betweenness, edge-cluster-coefficient Next lecture: Mon, Nov 4, 2013 • Modular decomposition • Robustness Short Test #1: Fri, Nov. 8